BZ
Back to side projects

AI Engineer (Lizzy AI capstone)

Legal Contract Advisor: High-Precision RAG for Legal Q&A

Contract Q&A RAG system for Lizzy AI built to deliver high-precision answers on legal documents using semantic chunking, hybrid retrieval, and a fully evaluated RAG pipeline.

Project gallery

Use the side arrows to browse; click an image for full size.

87% relevance on contract analysis through optimized chunking and hybrid retrieval.

Feb 2024·1 min read
PythonLangChainWeaviateHugging FaceFastAPIReactHybrid Retrieval

Legal Contract Advisor: High-Precision RAG for Legal Q&A

87% relevance on contract analysis: a RAG system for Lizzy AI that answers questions about legal contracts with precision high enough to trust inside a legal workflow.

The problem

Generic RAG over legal contracts returns plausible-but-wrong answers: naive chunking splits clauses mid-thought, and a single retrieval strategy misses either exact legal terms or semantic matches. In legal Q&A a confident wrong answer is worse than no answer.

What I built

  • Semantic + structural chunking that keeps clauses intact instead of cutting on fixed token windows.
  • Hybrid retrieval (dense embeddings + keyword) over a Weaviate vector store, so both exact legal terms and paraphrased questions resolve.
  • A FastAPI service + React UI for asking questions against an uploaded contract.
  • An evaluation loop that scored retrieval and answer relevance across a labelled question set.

How I measured quality

I built a labelled question/answer set over sample contracts and tracked answer relevance as I changed chunking and retrieval. Changes were kept only when the eval score moved, not on subjective spot-checks, and that iteration is how relevance reached 87%. The Medium write-up walks through each enhancement.

Stack

Python, LangChain, Weaviate, Hugging Face embeddings, FastAPI, React, hybrid retrieval.